The Week the Machines Got a Raise
A company that did not exist four years ago is now worth more than most countries. The machine it sells taught itself to rewrite sixteen thousand lines of code in a language it had never used. The people who used to write that code are looking for work. The machines themselves got forty-eight percent more expensive in two months. And a man who went to prison for money laundering is now richer than Bill Gates. It was a normal week.
Anthropic made one billion dollars a year in January of 2025. By April of 2026, it was making thirty billion. That is a thirty-fold increase in fifteen months. The company sells a machine that thinks, and the machine that thinks is now more popular than the one OpenAI sells, which is the machine that started the whole thinking-machine industry three and a half years ago. A thousand companies each pay Anthropic more than a million dollars a year for the privilege. Claude Code, the tool that writes software for other people, generates two and a half billion of those dollars all by itself. CFO Krishna Rao announced the thirty-billion number on April 6. The next day, Axios ran a headline that said, "No company in American history has ever grown like Anthropic." They are planning an IPO for October. The estimated valuation is somewhere between four hundred and five hundred billion dollars. For context, that is roughly the GDP of Austria. Austria has been around since 976 AD.
Here is what the machine can do now. Researchers at Epoch AI and METR built a test called MirrorCode. They gave an AI a program it had never seen and told it to build the whole thing again from scratch without looking at the original code. The AI could run the program and see what it did, but not how. Like being asked to rebuild a clock by listening to it tick. One of the programs was called gotree. It is a bioinformatics toolkit written in Go. Sixteen thousand nine hundred and five lines of code. Forty commands for parsing and manipulating phylogenetic trees. Four software engineers estimated it would take a skilled human two to seventeen weeks to reimplement it. Claude Opus 4.6 did it and passed two thousand of two thousand and one tests. It chose to rewrite the whole thing in Rust instead of Go, because nobody told it not to. The single test it failed involved a hidden date annotation that was not documented anywhere. A year earlier, the previous version of the same model managed fifteen percent. Two hundred and eighty million tokens. That is how many words the machine needed to think to itself before it finished.
In the first three months of 2026, seventy-eight thousand five hundred and fifty-seven people in the technology industry lost their jobs. Nearly half of those losses were attributed to artificial intelligence and automation. Entry-level hiring fell twenty-five percent. At major technology companies, only seven percent of new hires are recent graduates, down from nine point three percent in 2023. And yet. The Bureau of Labor Statistics reported that software developer employment grew three point eight percent in 2025. Global developer headcount reached twenty-eight point seven million, a new high. More developers than ever. Fewer jobs than expected. More code being written than at any point in human history. The math is not complicated. It is just unpleasant.
The machines are getting more expensive. A single Nvidia Blackwell GPU now costs four dollars and eight cents per hour to rent. Two months ago it cost two dollars and seventy-five cents. That is a forty-eight percent increase, tracked by the Ornn Compute Price Index, which was added to the Bloomberg Terminal on April 2 like any other commodity -- corn, crude oil, thinking machines. Ornn was founded by Kush Bavaria and Wayne Nelms, a former equity options trader at Susquehanna. They are now building futures contracts for GPU prices. You can hedge your exposure to the cost of artificial thought the same way a farmer hedges soybeans. The price is rising because the new AI systems do not just answer questions. They plan. They take actions. They spawn sub-processes and run for hours. They are hungrier than their predecessors and there are not enough GPUs on the planet to feed them all. Make the machines more efficient and people use more of them, not less. Jevons noticed this in 1865 with coal. He would have understood GPUs immediately.
Changpeng Zhao went to prison for four months. He had pleaded guilty in 2023 to enabling money laundering at Binance, the cryptocurrency exchange he founded, which at the time handled thirty-eight percent of all crypto trading on earth. The settlement cost Binance four point three billion dollars. Zhao personally paid fifty million. He was released on September 27, 2024. On October 23, 2025, President Trump signed a full pardon. In March of 2026, Forbes estimated Zhao's net worth at one hundred and ten billion dollars. Forty-seven billion more than the previous year. Two billion more than Bill Gates. The calculation was based on his estimated ninety percent stake in Binance, which Forbes valued using revenue multiples from publicly traded competitors like Coinbase. Zhao posted on X that the number was "definitely not accurate" and called the methodology "a guess a number exercise." Senator Elizabeth Warren noted that he had pleaded guilty to money laundering and then boosted Trump's cryptocurrency ventures. She did not say "so it goes," but she could have. He did not write any code.